R∞N — RAIN v6

AI Audio Transformation, Mastering & Distribution Infrastructure

ThatGuy Productions · ARCOVEL Technologies International

"Rain doesn't live in the cloud." — The render engine runs on your machine. Audio never leaves your device during processing.

RAIN is a full-stack, local-first AI mastering and distribution engine that brings studio-grade audio processing to independent artists and labels. It combines a deterministic C++/WASM DSP render engine with cloud-based AI inference, cryptographic provenance certificates, and direct streaming platform distribution — all in one product.

Beyond LANDR. Beyond iZotope. Beyond anything that came before.

---

Architecture

┌───────────────────────────────────────────────────────────────────────────────────┐
│  Browser  (React 19 + Vite 7 + TypeScript 5.5 + Tailwind 4)        │
│  ┌──────────────────┐  ┌─────────────────┐  ┌──────────────────────┐ │
│  │ Web Audio API    │  │ ONNX Runtime Web│  │ RainDSP  C++/WASM  │ │
│  │ 32-bit preview   │  │ WebGPU → WASM   │  │ 64-bit · deterministic│ │
│  └───────────────────┘  └─────────────────┘  └───────────────────────┘ │
└────────────────────────────────────┬──────────────────────────────────────┘
                                 │ REST / WebSocket  (FastAPI)
┌────────────────────────────────────▼──────────────────────────────────────┐
│  Backend  (Python 3.12 · FastAPI · Celery · Structlog)              │
│  ┌──────────────┐ ┌──────────────┐ ┌─────────────┐ ┌──────────────┐  │
│  │ Master Engine│ │ QC Engine  │ │ RAIN-CERT   │ │ DDEX ERN 4.3 │  │
│  │ 16-stage DSP │ │ 18 checks  │ │ Ed25519     │ │ AI disclosure │  │
│  └──────────────┘ └──────────────┘ └─────────────┘ └──────────────┘  │
│  ┌──────────────┐ ┌─────────────┐ ┌─────────────┐ ┌──────────────┐  │
│  │ Feature Ext. │ │ Heuristic  │ │ LabelGrid   │ │ Stripe       │  │
│  │ 43-dim vector│ │ 46 params  │ │ Distribution│ │ Billing      │  │
│  └──────────────┘ └─────────────┘ └─────────────┘ └──────────────┘  │
└─────────────────────────────────────────────────────────────────────────────┘
          │                      │                     │
   PostgreSQL 18          Valkey 9.0            MinIO / S3
   (RLS enabled)          (BSD-3 Redis fork)    (paid tiers only)

Dual-Path Design

Path	Engine	Precision	Purpose
Preview	Web Audio API	32-bit float	Real-time monitoring, < 50 ms latency
Render	RainDSP WASM	64-bit double	Deterministic, authoritative output

Determinism guarantee: Same input + same params + same WASM binary = bit-identical output, every time.

---

16-Stage Mastering Pipeline

Stage	Name	Description
01	Format Normalization	Resample to 48 kHz, 64-bit float stereo
02	Provenance Record	Ed25519 input hash, C2PA manifest init, AudioSeal seed
03	Feature Extraction	43-dim vector — Loudness (5), Dynamics (6), Spectral (16), Stereo (7), Transient (5), Tonal (4)
04	AI Inference	RainNet v2 → 46 ProcessingParams via sigmoid × 10 macro mapping
05	Reference Matching	Genre-aware spectral target matching
06	Spectral Repair	HPF, sibilance reduction, rumble removal, spectral smoothing
07	Source Separation	BS-RoFormer 4-pass cascade → 12 stems
08	Per-Stem Repair	Individual stem QC and spectral correction
09	Per-Stem Processing	SAIL v2 stem-aware limiting, vocal protection, gain faders
10	Master Bus	EQ → Multiband compression → Stereo widening → Groove → Life injection
11	Loudness Targeting	27 platform targets — Spotify −14, Apple −16, Atmos −18, CD −9, vinyl…
12	Spatial Rendering	Dolby Atmos HRTF binaural, M/S stereo enhancement
13	QC Validation	18 automated checks with auto-remediation
14	Forensic Watermark	16-bit AudioSeal, Chromaprint fingerprint
15	Output Packaging	24-bit WAV @ 48 kHz + 320 kbps MP3 with TPDF dither; RAIN-CERT signed
16	Distribution	DDEX ERN 4.3.2, LabelGrid API delivery, ISRC/UPC generation

---

7 Macro Controls

Emotionally-resonant, non-technical controls mapping to bounded subsets of the 46 DSP parameters. RainNet v2 outputs all 7 macros at indices 39–45 via sigmoid × 10 → [0.0, 10.0].

Macro	DSP Mapping
BRIGHTEN	High-shelf at 8 kHz + air peak at 16 kHz · 0 → +4 dB
GLUE	Multiband compression ratios/thresholds · 0 = transparent, 10 = bus glue 4:1
WIDTH	M/S side-channel gain · bass mono below 200 Hz enforced regardless
PUNCH	Mid-band transient shaping via attack/release · snare, kick, vocal presence
WARMTH	Low-shelf at 200 Hz + analog saturation · 0 = clean, 10 = +3 dB + tube sat
SPACE	Stereo decorrelation and M/S balance for depth · interacts with WIDTH
REPAIR	Spectral repair intensity, HPF, de-essing, noise floor

---

12-Stem Source Separation

Protect your authentic human vocal recordings. Upload AI-reworked structures, separate the stems, and cleanly replace AI vocals with your original takes.

Pass	Model	Output
1	BS-RoFormer SW	vocals · drums · bass · guitar · piano · other
2	MVSep Karaoke MelBand RoFormer	lead vocals + backing vocals
3	Spectral band-split (LarsNet pending)	kick · snare · hats · percussion
4	anvuew dereverb MelBand RoFormer	room/ambience + dry FX

Per-stem gain faders, solo/mute, 12-stem waveform display. SAIL v2 (Stem-Aware Intelligent Limiting) with dedicated vocal protection.

---

Provenance & Compliance

Standard	Status	Notes
EU AI Act Article 50	✅ Active	C2PA v2.2 + DDEX AI disclosure · deadline 2026-08-02
DDEX ERN 4.3.2	✅ Active	Full AI involvement fields per September 2025 standard
C2PA v2.2	✅ Active	CBOR-encoded manifests, Ed25519 signed
RAIN-CERT	✅ Active	Ed25519 provenance certificates on every render
AudioSeal	✅ Active	16-bit invisible watermarks (Meta, MIT licence)
Chromaprint	✅ Active	Audio fingerprints in PostgreSQL
ISO 3901 (ISRC)	✅ Active	Generated per standard with registrant code
AES17 True Peak	✅ Active	4× oversampling limiter · −1.0 dBTP ceiling

---

AI Co-Master Engineer

Claude Sonnet integration via claude_service.py (model configurable; default claude-sonnet-4-20250514):

• Natural language Intent Engine maps requests to bounded ProcessingParams deltas
• 7 macro suggestions with confidence scores (JSON-validated, clamped 0–10)
• Before/after mastering reports in plain language
• Tension-pair conflict detection (e.g. BRIGHTEN + WARMTH)
• Voice control via Web Speech API for hands-free mastering

Artist Identity Engine (AIE): 64-dimensional voice vector. Adaptive EMA (α = 0.90 stable, 0.60 cold-start). Personalizes after 5 sessions. Exportable as HMAC-SHA256 signed JSON.

---

Tech Stack

Layer	Technology
Frontend	React 19 · Vite 7 · TypeScript 5.5 · Tailwind 4 · Framer Motion 11
State	Zustand 5 · TanStack Query 5
Render Engine	RainDSP (C++20 / WASM via Emscripten) · 64-bit · deterministic
ML Inference	ONNX Runtime Web 1.24 (WebGPU → WASM fallback) · RainNet v2
Backend API	FastAPI 0.109+ · Python 3.12 · Celery · Structlog · SlowAPI
Database	PostgreSQL 18 with Row-Level Security on all tables
Cache / Queue	Valkey 9.0 (BSD-3-Clause Redis fork)
Object Storage	S3-compatible (MinIO in dev) · free tier = WASM-only, no S3
Provenance	Ed25519 · C2PA v2.2 · AudioSeal · Chromaprint · CBOR (RFC 8949)
Separation	BS-RoFormer SW · MelBand RoFormer · auto-download via pip
Distribution	DDEX ERN 4.3 · LabelGrid API · ISRC / UPC-EAN-13
Desktop / Plugin	Tauri 2.0 · JUCE 8 (VST3 / AU / AAX)

---

Getting Started

Prerequisites

• Docker Desktop 4.x+
• Node.js 20+
• Python 3.12+

Docker (recommended)

# Clone
git clone https://github.com/ThatGuy-Productions/RAIN-V6-AI-AUDIO-TRANSFORMATION-MASTERING-AND-DISTRIBUTION-INFRUSTRUCTURE.git
cd RAIN-V6-AI-AUDIO-TRANSFORMATION-MASTERING-AND-DISTRIBUTION-INFRUSTRUCTURE

# Configure environment
cp .env.example .env
# Edit .env — set ANTHROPIC_API_KEY, STRIPE_SECRET_KEY, LABELGRID_API_KEY, etc.

# Start full stack (PostgreSQL 18 + Valkey 9.0 + backend + frontend)
docker compose up --build -d

# Verify
curl http://localhost:8000/health

Local development

# Backend
cd backend
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

# Frontend (separate terminal)
cd frontend
npm install
npm run dev

GPU worker (stem separation)

# Requires CUDA 12+ and SEPARATION_ENABLED=true in .env
docker compose -f docker-compose.gpu.yml up -d worker

---

API Routes

The backend exposes 18 mounted API routers under /api/v1/.

Router	Prefix	Auth	Description
auth.py	/auth	Public / JWT	Registration, login, JWT refresh, password reset
upload.py	/upload	JWT	Audio file upload, S3 pre-signed URLs, file validation
master.py	/master	JWT	Session creation, mastering pipeline trigger, status polling, parameter overrides
separate.py	/separate	JWT · Creator+	BS-RoFormer 12-stem separation jobs, stem download
download.py	/download	JWT	Signed download URLs for rendered masters and stems
distribution.py	/distribution	JWT · Producer+	DDEX ERN 4.3.2 package generation, LabelGrid delivery, ISRC/UPC issuance
billing.py	/billing	JWT	Stripe checkout sessions, subscription management, webhook receiver
aie.py	/aie	JWT · Artist+	Artist Identity Engine — vector query, export, reset
sessions.py	/sessions	JWT	Session history, status, output metadata, RAIN Score retrieval
qc.py	/qc	JWT	Manual QC re-run, 18-point check report
provenance_routes.py	/provenance	JWT	RAIN-CERT retrieval, C2PA manifest query, Ed25519 verification
waitlist.py	/waitlist	Public	Beta waitlist registration
assist.py	/assist	JWT	Claude AI assistant — intent parsing, macro suggestions, mastering advice, voice command processing
lora.py	/lora	JWT · Enterprise	Custom LoRA adapter training jobs, model version management, inference switching
suno_import.py	/suno	JWT	Suno AI track import — fetch by song ID or URL, normalize, queue for mastering
whitelabel.py	/whitelabel	JWT · Enterprise	White-label API key provisioning, partner branding config, usage metering
workspaces.py	/workspaces	JWT · Studio+	Multi-artist workspace management, collaborator roles, shared session access
score.py	/score	Public (10 req/hr/IP)	Public RAIN Score endpoint — no auth required, returns 0-100 composite quality metric

Auth tiers: Public = no token required · JWT = any authenticated user · tier suffixes (Creator+, Artist+, etc.) = minimum subscription required.

---

Pricing

Tier	Price	Renders / mo	Key Features
Casual	Free	Preview only	Real-time WASM mastering preview, RAIN Score, no file export
Creator	$9	50 downloads	Full-res export, WAV / FLAC / MP3, Simple Mode
Independent Artist	$29	10 renders	Stem separation, Claude AI (10/mo), Artist Identity Engine
Producer	$59	25 renders	DAW plugin, Distribution Intelligence, RAIN-CERT
Studio	$149	75 renders	Dolby Atmos, DDEX / DDP, vinyl mastering, collaboration
Label / Distributor	$349	300 renders	Multi-artist roster, batch processing, LabelGrid direct
Enterprise	Custom	Unlimited	Custom RainNet LoRA, white-label API, dedicated support

Annual billing: ~20% discount. Contact philippusbolke@gmail.com for enterprise licensing.

---

Architecture Rules

These constraints are immutable. See CLAUDE.md for the full specification.

1. Local-First Processing — RainDSP WASM is the sole render engine. Audio never reaches S3 on the free path. Free renders live in WASM memory and are discarded on session close.

2. Dual-Path Architecture — Preview (Web Audio API, 32-bit) and Render (RainDSP, 64-bit) are always separate codepaths. Never merge them.

3. Multi-Tenant Isolation — Every DB query includes WHERE user_id = $user_id. Row-Level Security enabled on all tables. No exceptions.

4. K-Weighting Sign Convention — y = b0·x + b1·x₁ + b2·x₂ − a1·y₁ − a2·y₂. a1 stored negative, subtracted. ITU-R BS.1770-4 convention — never change this.

5. NORMALIZATION_VALIDATED Gate — RAIN_NORMALIZATION_VALIDATED=true → gate OPEN → RainNet inference active. Closed by default. Sign-off authority: Phil Bölke.

6. WASM Binary Integrity — rain_dsp_wasm_hash verified at session start via SHA-256. Mismatch = RAIN-E304, render blocked. Set RAIN_EXPECTED_WASM_HASH in production.

7. Provenance Before Output — Output hash verified against RAIN-CERT before session marked complete. RAIN-E305 on mismatch. RAIN-E306 on unsigned cert. No exceptions in production.

---

Project Structure

RAIN-V6-AI-AUDIO-TRANSFORMATION-MASTERING-AND-DISTRIBUTION-INFRUSTRUCTURE/
├── backend/                   FastAPI application
│   ├── app/
│   │   ├── api/routes/        19 routers — see API Routes table above
│   │   ├── core/              config, database, security, observability, audio_io
│   │   ├── models/            SQLAlchemy ORM models (PostgreSQL 18 + RLS)
│   │   ├── schemas/           Pydantic request / response schemas
│   │   └── services/          master_engine, qc_engine, rain_score_v2, separation…
│   └── tests/                 pytest async test suite
├── frontend/                  React 19 SPA (Vite 7 · TypeScript · Tailwind 4)
├── rain-dsp/                  C++20 DSP engine — Emscripten WASM build
├── rain-desktop/              Tauri 2.0 desktop wrapper
├── rain-plugin/               JUCE 8 VST3 / AU / AAX plugin
├── ml/                        PyTorch training, ONNX export, RainNet v2
├── docker/                    Dockerfiles (backend, frontend, gpu-worker)
├── monitoring/                Prometheus + Grafana dashboards
├── docs/                      Architecture docs, pipeline specs
├── CLAUDE.md                  Immutable architecture specification
├── RAIN-BLUEPRINT.md          Full technical blueprint
└── .env.example               All required environment variables documented

---

Error Code Reference

Code	Meaning	Action
RAIN-E100	JWT authentication failure	Re-authenticate; check token expiry
RAIN-E304	WASM binary hash mismatch	Set RAIN_EXPECTED_WASM_HASH to deployed hash
RAIN-E305	Output hash mismatch in RAIN-CERT	Re-render; check CERT signing key
RAIN-E306	Session completed without signed cert	Check C2PA key paths in config
RAIN-E503	Public score rate limit exceeded	10 requests/hour per IP — wait or authenticate
RAIN-E620	Separation disabled	Set SEPARATION_ENABLED=true and provision model checkpoints
RAIN-E621	demix() not available	Install music-source-separation-training on GPU worker
RAIN-E900	Claude API auth failure	Check ANTHROPIC_API_KEY in env
RAIN-E901	Claude API timeout / exhausted retries	3 retries with exponential backoff — check network
RAIN-E902	Claude response parse failure	Invalid JSON from model — check logs for raw response

---

License

Proprietary — © 2026 ThatGuy Productions. All rights reserved.

Contact philippusbolke@gmail.com for licensing enquiries.

---

RAIN v6 · Engine stamp: RAIN v6 — BS-RoFormer 12-stem · LLM advisory layer: configurable Claude Sonnet (claude-sonnet-4-20250514 default) · Publisher: ThatGuy Productions